Re: [Qemu-devel] [RFC PATCH v2 1/4] net: port tap onto glib

[Stefan Hajnoczi <stefanha@gmail.com>]

On Tue, Apr 09, 2013 at 01:12:39PM +0800, liu ping fan wrote:
> On Mon, Apr 8, 2013 at 7:44 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> > On Wed, Apr 03, 2013 at 05:28:39PM +0800, liu ping fan wrote:
> >> On Thu, Mar 28, 2013 at 10:32 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> >> > On Thu, Mar 28, 2013 at 03:55:52PM +0800, Liu Ping Fan wrote:
> >> >> From: Liu Ping Fan <pingfank@linux.vnet.ibm.com>
> >> >>
> >> >> Bind each NetClientState with a GSource(ie,NetClientSource). Currently,
> >> >> these GSource attached with default context, but in future, after
> >> >> resolving the race between handlers and the interface exposed by NetClientInfo
> >> >> and other re-entrant issue,  we can run NetClientState on different threads
> >> >>
> >> >> Signed-off-by: Liu Ping Fan <pingfank@linux.vnet.ibm.com>
> >> >> ---
> >> >>  include/net/net.h |   27 +++++++++++++++
> >> >>  net/net.c         |   96 +++++++++++++++++++++++++++++++++++++++++++++++++++++
> >> >>  net/tap.c         |   57 +++++++++++++++++++++++++------
> >> >>  3 files changed, 169 insertions(+), 11 deletions(-)
> >> >
> >> > Please split this into two patches:
> >> >
> >> > 1. NetClientSource
> >> > 2. Convert tap to NetClientSource
> >> >
> >> > Once you do that it turns out that NetClientSource has nothing to do
> >> > with the net subsystem, it's a generic file descriptor GSource (weird
> >> > that glib doesn't already provide this abstraction).
> >> >
> >> > Each net client needs to reimplement .bind_ctx() anyway, so I don't see
> >> > much point in having NetClientSource.nsrc[].  We might as well let net
> >> > clients have that field themselves and destroy the GSource in their
> >> > destructor function.
> >> >
> >> The only way to detach the GSource from GMainContext is
> >> g_source_destroy, so if we want to re-bind nc from threadA to threadB,
> >>  we should destroy the old one and create a new. Is that meaningful?
> >
> > I guess that can be done.
> >
> > What I was really thinking when I suggested getting rid of nsrc[] is
> > that it's a little ugly to have the array with 2 GSources.  Different
> > net clients have different numbers of GSources - 0 for NICs, 1 for most
> > backends, 2 for ioeventfd for virtio-net data plane with separate rx/tx.
> >
> > So my thought was to leave the number of GSources in the layer that uses
> > them - each specific net client.
> >
> OK, see. What about nsrc[0] at the end of struct or **nsrc ?

The net core code doesn't need to know about nsrc at all.  I think
nsrc[] should be dropped completely.  The net backends should manage
their GSource (if they need one).

Stefan